1 / 11

Problem: General Name Transcription Improvement

Improved Name Recognition with Meta-data Dependent Name Networks published by Sameer R. Maskey, Michiel Bacchiani, Brian Roark, and Richard Sproat presented by Irina Likhtina. Problem: General Name Transcription Improvement. Solutions that use no prior knowledge require large increase in

scot
Download Presentation

Problem: General Name Transcription Improvement

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Improved Name Recognitionwith Meta-data Dependent Name Networkspublished by Sameer R. Maskey, Michiel Bacchiani,Brian Roark, and Richard Sproatpresented by Irina Likhtina

  2. Problem: General Name Transcription Improvement Solutions that use no prior knowledge require large increase in • Size • Complexity • Ambiguity because any improvement accomplished before was possible only through an increase in lexicon used.

  3. Paper’s Solution • Use meta-data at runtime in the form of • Caller ID string – “cname” • Name of mailbox owner – “mname” This meta-data is reasonably available due to the prevalence of caller identification given by phone companies.

  4. Database Used • Scanmail training corpus of 100 hours of voicemail messages from 140 employees of AT&T. • Manually transcribed with “cname” and “mname” tags • Gender balanced • ~12% non-native speakers • 238 random messages for testing, everything else (~ 10,000 messages) for training

  5. Approach • Three steps of the algorithm: • create a class-based language model • create a name network that will give instances for the classes of the model • replace the class-based language model at runtime with the name networks

  6. Class-Based Language Model • Manual tags of mailbox name and caller name for each message replaced with “mname” and “cname” labels • “mname” and “cname” represented the 2 class tokens that can be substituted with any values in the future

  7. Name Network • To get the values for “mname” and “cname”, an internal AT&T employee directory (~ 40,000 people) listing was used • “cname” created from variations of static titles (Miss, Mr), full first names and nicknames (Alexander, Alex), and last names (Jones)

  8. Name Network (continued) • Probability within class – training corpus • Probability within first names – AT&T directory listing

  9. Replacement in the ASR network • State of art before proposed algorithm: off-line composition, determinization and optimization of one existing grammar G and one lexicon L Proposed algorithm is impracticalwith off-line composition because of so many variations needed • Proposed algorithm: L by G optimization done for each class using G specific for that class Small overhead of proposed algorithm compared to off-line optimization

  10. Experimental results • Word Error Rates (WER) improvement small • Absolute reduction of 0.6% • Named Error Rate (NER) improvement significant • Absolute reduction of 20 %

  11. Conclusions • Large reduction in NER is very critical: • Name transcription is the goal • Scanmail users expressed a strong desire for the system to recognize names correctly • Other improvements: • Errors in OOV • In-vocabulary name recognition • No significant increase in complexity with good name coverage • No need for manual design when the system is moved to new environment

More Related